I want to analyze experimental data with a linear mixed-effects model (in R). I have an hypothesis about one fixed factor, but I also have many random factors and continuous covariates that can potentially affect the results (subject, gender, age, item, item frequency, etc.).
I tried constructing the best fitting model, but I couldn't converge on one model.
I added variables to the model one-by-one (and kept only those that yielded a significant result on the anova comparing the new model with its predecessor). Then, I ran the step function on the final model (backward elimination of parameters, from the lmerTest package), and it threw away several of those variables that seemed to make a significant contribution to the model.
Any suggestions on how to approach this model selection problem?
Thanks,
Chen
Thank you, Mercedes.
I like your suggestion. I assume it works only for non-categorical variables, right?
Chen
Look carefully at the attached document and decide which type of model you are trying to build. If the model is predictive, I recommend lasso and have attached an additional paper giving examples of lasso selection as well as gradient boosting. I hope that you find this useful. Best regards, David Booth
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA